Andrej Karpathy • 12/19/2025

2025 LLM Year in Review

This article reviews major developments in Large Language Models (LLMs) in 2025. It highlights the rise of Reinforcement Learning from Verifiable Rewards (RLVR) as a core training stage, enabling models to develop reasoning strategies. It also discusses a conceptual shift in understanding LLM intelligence, moving from an 'animal' to a 'ghost' metaphor, reflecting their unique and non-biological nature.

0 comments

#AI Research #Reinforcement Learning #LLM Training